Re-extraction after manual tool causes missing unet-dfl masks

[Surrogator]

1. I've extracted faces with a unet-dfl masker.

2. I've sorted faced and removed false positives and some misaligned extractions.

3. I've used the manual tool to find the missing faces and set alignments for them.

4. I've deleted the original extracted faces and re-extracted the faces using the new alignments file.

5. Upon converting the video, I get many errors like "Failed to convert image: 'vid_007383.png'. Reason: 'unet-dfl'"

How can I re-extract the faces including the unet-dfl mask for these new alignments?

[torzdf]

When you manually fix the mask, any generated masks are erased (they have to be, as the masks no longer correspond to the actual face).

It's not ideal, but you can regenerate masks with the "mask" tool. If you select the mask(s) you want, select "missing" then it will fill in the gaps.

[kellurian]

When I extract a data set, I typically use vgg_obstructed. However, every time I try to do a manual edit of theses, invariably when I go to train I get the error that the mask I have been training on is not present . This even happens when I do a mask edit and use the vgg_obstructed mask view. Is it defaulting to components and extended, or should I just use those from now on? Is there a way to make sure that all of the masks and faces are saved to a vgg mask before saving the alignment file? or should I not worry about training on the new alignment file, only train on the original alignment and just use the new one(after manual alignments) for the convert?

09/27/2020 14:39:26 ERROR Caught exception in thread: '_training_0'
09/27/2020 14:39:26 ERROR At least one of your faces does not have the mask vgg-obstructed stored for it.
09/27/2020 14:39:26 ERROR You should run the Mask Tool to generate this mask for your faceset or select a different mask in the training configuration options.
09/27/2020 14:39:26 ERROR The face that caused this failure was [side: A, frame: combined_002186.png, index: 0]. The masks that exist for this face are: ['components', 'extended'].
09/27/2020 14:39:26 ERROR Be aware that there are probably more faces without this Mask Type

[bryanlyon]

Chances are you are changing some faces/alignment and those masks are being reset since the face image has changed. Don't do combined edit alignments and masks in the same pass, do separate passes for each.

[kellurian]

Ok, well I tried that. Only did manual with the alignments, didn't touch the masks at all and still getting the same error. Could this be something with the vgg obstructed masks? like does the manual tool just erase all those masks when you start using alignment edits as some kind of default?

[torzdf]

The reason for this is simple, but slightly annoying.

When you edit landmarks, the mask for the face changes as well. This is fine for components + extended, as they are landmarks based, so they get auto-regenerated. The other masks are NN based, It's not possible to regenerate these masks on the fly as the GPU is being used for Face Alignment, so the corresponding masks are deleted on any edit.

Once you have finished manually editing the landmarks, use the mask tool to regenerate the missing masks for your newly corrected faces. You can then go in and edit them.